The Research Of SAR Ship Target Recognition Based On Deep Learning

Because of the advantages of adaptive capacity to all weather and all time,synthetic aperture radar(SAR)is widely used on many fields such as battlefield scout,marine surveillance,imaging guidance,damage assessment,et al.Nowadays,traditional algorithms domain the SAR ship target recognition,but still have problem with practical application,due to the weak capability to capture more various and deeper features.Own to the enhancement of hardware compute ability and the increase of the amount of data,it has been promoted in recent years that the ability of detection and recognition using deep learning technology,and has replaced the traditional algorithms.This thesis introduce the recent research finding in deep learning into the field of SAR ship target recognition,and research on next 4 parts:1?In the first place,the convolutional neural networks based on the region proposal are researched to be used to detect and recognize microscale ship targets in SAR images.The frameworks of Faster R-CNN and R-FCN are analyzed,and the improvement measure is proposed to adapt the SAR ships images and the microscale.The improved algorithm is implicated with 3 kinds of network models including ZF-net,VGG-16,ResNet-50,and the results of corresponding experiments are provided.2?Because of the large data demand and the difficulty of acquiring samples of non-cooperative ship targets,a method is proposed that using generative adversarial networks(GAN)to augment the amount of samples.Deep convolutional GAN is adopted to generate verisimilitude SAR ship chips.An improved measurement is proposed to reduce the amount of samples required,and achieve success with only 250 ship chips.3?A method is proposed to detect and recognize ship targets in SAR images with limited samples,which use modified Faster R-CNN or R-FCN to detect and recognize SAR ship targets,and the deep convolutional generative adversarial network is used to augment the training samples of Faster R-CNN and R-FCN.The experiments verify the feasibility of this method with limited samples.4?According to the methods proposed in this thesis,a SAR ship target recognition algorithm training system has been designed and developed,on the purpose of reducing the difficulty of operation during practical application.By encapsulate the algorithms and graphical user interface,this system not only realizes utility functions concluding both training and testing,but also provides proposed methods with a way to engineering application,which has a friendly interface and is easy to operate.